Irfanuruchi/building-engineering-synthetic-dataset-v5

Building Engineering Synthetic Dataset (V5)

Repository: Irfanuruchi/building-engineering-synthetic-dataset-v5

This repository contains a synthetic dataset for training engineering reasoning models focused on building engineering calculations and sanity checks.

The dataset was generated using physics-based engineering equations and structured prompts suitable for LLM fine-tuning.

It was used to train:

Irfanuruchi/qwen2.5-1.5b-buildeng-precheck-lora-v5

---

Dataset Overview

The dataset contains engineering problems covering:

• HVAC airflow sizing
• Hydronic pipe sizing
• Fan power calculations
• Pump power calculations
• Cooling coil loads
• Heat transfer calculations
• Structural engineering pre-checks
• Engineering sanity checks
• Unit consistency traps
• Missing input detection

The dataset is designed to teach models to:

• perform engineering calculations
• show reasoning steps
• detect unrealistic assumptions
• refuse incomplete design problems

---

Dataset Size

Split	Samples
Train	~57,000
Validation	~3,000
Total	~60,000

---

Data Format

Each dataset row is stored as JSONL with the following format:

{
  "text": "<|system|>\nYou are a building engineering assistant.\n<|user|>\nSize a round air duct.\nAirflow=3500 m³/h\nVelocity=6 m/s\n<|assistant|>\nSelect round duct ≈ Ø480 mm\n..."
}

The dataset follows the Qwen chat format:

<|system|>
<|user|>
<|assistant|>

This allows direct use for instruction tuning.

---

Engineering Categories

HVAC

• duct sizing
• airflow calculations
• fan power estimation

Hydronics

• pipe velocity sizing
• water flow calculations
• pump power estimation

Heat Transfer

• wall heat loss estimation

Structural Engineering

• beam bending stress
• beam deflection estimation
• tributary load calculation
• slab load estimation

---

Engineering Sanity Checks

A significant portion of the dataset teaches the model to reject unrealistic inputs.

Examples include:

Unit traps

Example:

Airflow = 0.3 m³/h
Velocity = 5 m/s

The model should recognize this airflow is physically unrealistic for HVAC systems.

---

Velocity traps

Example:

Velocity = 15 m/s in a quiet space

The model should reject this assumption because typical velocities are 3–6 m/s.

---

ΔT traps

Example:

Water ΔT = 1 K

The model should detect that this results in unrealistically high flow rates.

---

Missing inputs

Example:

Size supply duct for conference hall

The model should respond that airflow is required before sizing.

---

Dataset Generation

The dataset is generated using a Python script that uses engineering equations to produce valid examples.

Examples include:

• duct diameter from airflow and velocity
• pump power from flow and head
• cooling load from airflow and temperature difference
• beam stress from moment and section modulus

This ensures the dataset is numerically consistent and physically plausible.

---

File Structure

train.jsonl
valid.jsonl
make_dataset_v5.py
README.md

---

Intended Use

This dataset is designed for:

• training engineering reasoning models
• fine-tuning LLMs for technical domains
• research in domain-specialized language models

It is particularly suited for LoRA fine-tuning of small and medium LLMs.

---

Limitations

This dataset focuses on simplified engineering problems.

Limitations include:

• simplified structural models
• simplified HVAC assumptions
• no full engineering standards implementation

Results produced by models trained on this dataset must always be verified using professional engineering standards (ASHRAE, Eurocode, etc.).

---

Future Improvements

Future dataset versions may include:

• larger dataset sizes (>100k problems)
• more complex structural engineering tasks
• multi-step engineering reasoning chains
• additional HVAC system design constraints

---

Author

Irfan Uruchi

This dataset is part of research into lightweight domain-specialized LLMs for engineering reasoning(also used as test for the 32B LLM model im fine tuning for building engineering).